| commit | 560e6af7b1f74e9c020a0f82817f9d926e0c4f72 | [log] [tgz] |
|---|---|---|
| author | Harshit Aghera <haghera@nvidia.com> | Mon Apr 21 20:04:56 2025 +0530 |
| committer | Ed Tanous <ed@tanous.net> | Fri Jun 06 20:47:25 2025 +0000 |
| tree | fd847f6ed009220327e016fd53caa4c4693a4a46 | |
| parent | 490fc51e44727704e4402cf94fc20a162e6c09c2 [diff] |
nvidia-gpu: add support for communication to the endpoint The commit uses MCTP VDM protocol to read temperature sensor value from the gpu. The MCTP VDM protocol is an extension of the OCP Accelerator Management Interface specification. [1] Tested: Build an image for gb200nvl-obmc machine with the following patches cherry picked. This patches are needed to enable the mctp stack. https://gerrit.openbmc.org/c/openbmc/openbmc/+/79422 Restart the nvidiagpusensor service. ``` root@gb200nvl-obmc:~# systemctl start xyz.openbmc_project.nvidiagpusensor.service ``` The app is detecting entity-manager configuration on gb200nvl-obmc machine. The app is also able to detect all the endpoints from the mctp service dbus tree. The app is reading temperature sensor value from gpu correctly and the temperature sensor is also present on redfish. ``` $ curl -k -u 'root:0penBmc' https://10.137.203.137/redfish/v1/Chassis/NVIDIA_GB200_1/Sensors/temperature_NVIDIA_GB200_GPU { "@odata.id": "/redfish/v1/Chassis/NVIDIA_GB200_1/Sensors/temperature_NVIDIA_GB200_GPU", "@odata.type": "#Sensor.v1_2_0.Sensor", "Id": "temperature_NVIDIA_GB200_GPU", "Name": "NVIDIA GB200 GPU", "Reading": 36.4375, "ReadingRangeMax": 127.0, "ReadingRangeMin": -128.0, "ReadingType": "Temperature", "ReadingUnits": "Cel", "Status": { "Health": "OK", "State": "Enabled" } }% root@gb200nvl-obmc:~# busctl tree xyz.openbmc_project.GpuSensor └─ /xyz └─ /xyz/openbmc_project └─ /xyz/openbmc_project/sensors └─ /xyz/openbmc_project/sensors/temperature └─ /xyz/openbmc_project/sensors/temperature/NVIDIA_GB200_GPU root@gb200nvl-obmc:~# busctl introspect xyz.openbmc_project.GpuSensor /xyz/openbmc_project/sensors/temperature/NVIDIA_GB200_GPU NAME TYPE SIGNATURE RESULT/VALUE FLAGS org.freedesktop.DBus.Introspectable interface - - - .Introspect method - s - org.freedesktop.DBus.Peer interface - - - .GetMachineId method - s - .Ping method - - - org.freedesktop.DBus.Properties interface - - - .Get method ss v - .GetAll method s a{sv} - .Set method ssv - - .PropertiesChanged signal sa{sv}as - - xyz.openbmc_project.Association.Definitions interface - - - .Associations property a(sss) 1 "chassis" "all_sensors" "/xyz/openbmc… emits-change xyz.openbmc_project.Sensor.Value interface - - - .MaxValue property d 127 emits-change .MinValue property d -128 emits-change .Unit property s "xyz.openbmc_project.Sensor.Value.Unit.… emits-change .Value property d 36.3125 emits-change writable xyz.openbmc_project.Sensor.ValueMutability interface - - - .Mutable property b true emits-change xyz.openbmc_project.State.Decorator.Availability interface - - - .Available property b true emits-change writable xyz.openbmc_project.State.Decorator.OperationalStatus interface - - - .Functional property b true emits-change ``` [1] https://www.opencompute.org/documents/ocp-gpu-accelerator-management-interfaces-v1-pdf Change-Id: Ied938b9e5c19751ee283b4b948e16c905c78fb48 Signed-off-by: Harshit Aghera <haghera@nvidia.com>
dbus-sensors is a collection of sensor applications that provide the xyz.openbmc_project.Sensor collection of interfaces. They read sensor values from hwmon, d-bus, or direct driver access to provide readings. Some advance non-sensor features such as fan presence, pwm control, and automatic cpu detection (x86) are also supported.
runtime re-configurable from d-bus (entity-manager or the like)
isolated: each sensor type is isolated into its own daemon, so a bug in one sensor is unlikely to affect another, and single sensor modifications are possible
async single-threaded: uses sdbusplus/asio bindings
multiple data inputs: hwmon, d-bus, direct driver access
A typical dbus-sensors object support the following dbus interfaces:
Path /xyz/openbmc_project/sensors/<type>/<sensor_name> Interfaces xyz.openbmc_project.Sensor.Value xyz.openbmc_project.Sensor.Threshold.Critical xyz.openbmc_project.Sensor.Threshold.Warning xyz.openbmc_project.State.Decorator.Availability xyz.openbmc_project.State.Decorator.OperationalStatus xyz.openbmc_project.Association.Definitions
Sensor interfaces collection are described here.
Consumer examples of these interfaces are Redfish, Phosphor-Pid-Control, IPMI SDR.
dbus-sensor daemons are reactors that dynamically create and update sensors configuration when system configuration gets updated.
Using asio timers and async calls, dbus-sensor daemons read sensor values and check thresholds periodically. PropertiesChanged signals will be broadcasted for other services to consume when value or threshold status change. OperationStatus is set to false if the sensor is determined to be faulty.
A simple sensor example can be found here.
Sensor devices are described using Exposes records in configuration file. Name and Type fields are required. Different sensor types have different fields. Refer to entity manager schema for complete list.
ADC sensors are sensors based on an Analog to Digital Converter. They are read via the Linux kernel Industrial I/O subsystem (IIO).
One of the more common use cases within OpenBMC is for reading these sensors from the ADC on the Aspeed ASTXX cards.
To utilize ADC sensors feature within OpenBMC you must first define and enable it within the kernel device tree.
When using a common OpenBMC device like the AST2600 you will find a "adc0" and "adc1" section in the aspeed-g6.dtsi file. These are disabled by default so in your system-specific dts you would enable and configure what you want with something like this:
iio-hwmon { compatible = "iio-hwmon"; io-channels = <&adc0 0>; ... } &adc0 { status = "okay"; ... }; &adc1 { status = "okay"; ... };
Note that this is not meant to be an exhaustive list on the nuances of configuring a device tree but really to point users in the general direction.
You will then create an entity-manager configuration file that is of type "ADC" A very simple example would like look this:
"Index": 0, "Name": "P12V", "PowerState": "Always", "ScaleFactor": 1.0, "Type": "ADC"
When your system is booted, a "in0_input" file will be created within the hwmon subsystem (/sys/class/hwmon/hwmonX). The adcsensor application will scan d-bus for any ADC entity-manager objects, look up their "Index" value, and try to match that with the hwmon inY_input files. When it finds a match it will create a d-bus sensor under the xyz.openbmc_project.ADCSensor service. The sensor will be periodically updated based on readings from the hwmon file.